History compare AR perf

Objective: Run and repair history-compare under Slate AR until target-backed evidence is green, plateaued, or blocked by correctness/architecture proof.

Completion threshold:

• Done when history-compare emits a benchmark-native primary METRIC line and one of these is true: two correctness-green repeat packets are at or below history_compare_worst_p95_ratio <= 2.0, two correctness-green packets plateau under 5% improvement, or a remaining optimization is blocked by concrete correctness or architecture evidence.

Verification surface:

• Target setup: pnpm bench:targets:check, pnpm bench:targets:dry-run -- history-compare, and target report refresh.
• Benchmark: HISTORY_BENCH_LEGACY_REPO=../../../slate bun run bench:history:compare:local in .tmp/slate-v2.
• Correctness: .tmp/slate-v2 bun check for evidence used to close the lane.

Constraints:

• Keep scope to the history compare target, benchmark script, target registry, AR session files, and generated target reports unless the benchmark exposes a real runtime owner.
• Do not commit, push, or open a PR unless explicitly requested.
• Prefer benchmark-native truth over wrapped wall-clock timing.

Boundaries:

• Source of truth: benchmarks/targets/slate-v2.json target history-compare and .tmp/slate-v2/scripts/benchmarks/core/compare/history.mjs.
• Browser surface: none; this is a Bun/jsdom-free history benchmark.
• Release artifact: slate-history patch changeset because runtime history replay behavior changed.

Blocked condition:

• Block only if the legacy repo cannot run, .tmp/slate-v2 correctness fails from an unrelated owner that cannot be isolated, or further improvement needs a public history API/runtime architecture decision.

Work Checklist:

• Inspect target registry and history benchmark script before edits.
• Upgrade or verify native metric output.
• Run target dry-run/check/report path.
• Run benchmark evidence and correctness proof.
• Record release, browser, PR, tracker, and review decisions.
• Run the mechanical autogoal completion check.

Phase / pass table:

Phase	Status	Evidence	Next
Intake	complete	target registry and history benchmark read	implementation done
Implementation	complete	native metrics, stable iteration count, and history replay optimization added	verification done
Verification	complete	AR runs 11 and 12 under 2.0, checks green	closeout done
Closeout	complete	plan updated and mechanical check passed	final response

Findings:

• The target was still using wrapped benchmark_seconds; that was not a useful history comparison metric.
• Initial native metric runs were red: default 2.63x, 15-iteration 3.1x, worst on typing undo.
• Generic transaction replay was cloning each historic operation and historic replay did not explicitly skip normalization like legacy Slate does.
• Adjacent text history operations can be compacted during undo/redo replay without changing the final document value.

Decisions and tradeoffs:

• Chose history_compare_worst_p95_ratio as the primary metric and kept mean ratio/delta as secondary evidence.
• Set the benchmark default to 15 iterations because 3-sample p95 is bullshit for sub-ms undo/redo lanes.
• Added a slate-history patch changeset because this is package runtime behavior, not just benchmark plumbing.
• No PR, commit, or tracker sync: not requested.
• Browser proof is N/A: this is a core/history benchmark and Bun test surface.

Verification evidence:

• node --check .tmp/slate-v2/scripts/benchmarks/core/compare/history.mjs passed.
• pnpm bench:targets:check passed.
• pnpm bench:targets:dry-run -- history-compare passed with metric=history_compare_worst_p95_ratio.
• pnpm bench:targets:report regenerated target history/report files.
• AR parser lint accepted the native history_compare_worst_p95_ratio metric.
• Targeted history tests passed: cd packages/slate-history && bun test ./test/history-contract.ts ./test/integrity-contract.ts ./test/document-state-history-contract.ts with 51 pass, 0 fail.
• Heavy guard run with HISTORY_BENCH_TYPE_OPS=200 measured history_compare_worst_p95_ratio=1.61.
• AR run 11 measured history_compare_worst_p95_ratio=1.9, history_compare_worst_mean_ratio=0.57; benchmark exited 0 and bun check passed.
• AR run 12 measured history_compare_worst_p95_ratio=0.38, history_compare_worst_mean_ratio=0.37; benchmark exited 0 and bun check passed.
• Final bash ./autoresearch.checks.sh passed in .tmp/slate-v2: lint had one non-blocking React Hook warning in site/examples/ts/pagination.tsx, typecheck passed, Bun tests 1172 pass, 95 skip, 0 fail, slate-layout 41 pass, slate-react Vitest 56 files, 590 tests passed.

Reboot status:

Question	Answer
Where am I?	Closeout complete
Where am I going?	Final response
What is the goal?	Make history-compare truthful and decide whether it needs optimization
What have I learned?	The red lane was real history replay overhead, then later p95 noise after the fix
What have I done?	Upgraded the metric contract, optimized historic replay, added a changeset, and proved two green AR packets

Open risks:

• Low: the p95 metric can still show GC spikes in fragment lanes, but both repeat packets stayed under target and mean ratios are now faster than legacy.
• Known non-blocking warning: site/examples/ts/pagination.tsx hook dependency warning appears in bun check but exits 0.